Eddy current pulsed thermography is a new nondestructive testing method that is widely used in metal structure testing. However, the extraction of features for crack detection and identification relies on manual experience, and the degree of automation and intelligence is insufficient. By combining the characteristics of eddy current pulsed thermography with a recurrent neural network (RNN), a bidirectional long short-term memory (Bi-LSTM)-based eddy current pulse thermography method is proposed for metal fatigue crack classification and recognition. The Bi-LSTM model was designed to enhance the transient information in the feature vectors. In the experiments, an eddy current heating device was used to heat the tested metal specimens. A real-time dataset was created using an infrared thermal camera that collected sequences of images. The Bi-LSTM model was trained on thermal images of cracks of different sizes and tested. Experimental analyses show that the Bi-LSTM network can be effectively applied for metal fatigue crack detection and recognition, with the detection accuracy reaching 100% for the cracks used in the experiments, which is superior to that of traditional neural networks and other deep learning models.